On October 12, 2026, Wang Xingxing, co-founder of Unitree Robotics, made a bold prediction in public: “Next year, GPT-7 may directly solve the embodied intelligence problem.” Compared to his earlier statement on August 20, 2026—“at the earliest two to three years, at the latest five to ten years”—the projected timeline for this industry inflection point has been significantly shortened to “as soon as one to two years.” This rapid revision is interpreted as a passive adjustment driven by the pace of iterations in general-purpose large models.
On the empirical front, the trend has shifted from simulation to real robots: RoboDojo’s September 2026 evaluation showed that GPT‑6 Astra achieved an average success rate of 22.48% across 42 dual‑arm tasks, surpassing the top scores of 40 publicly available specialized policies (19.34%); meanwhile, Artificial Analysis’s October 14 release of the AA‑Robotics v0.1 real‑robot benchmark further confirmed that, although Astra ranked first with a score of 69.4, its average time per task was approximately 30 minutes (compared to just 20 seconds for humans), with invocation costs as high as $16.36, and it suffers from fundamental execution flaws such as “inability to accurately determine task completion.”
Meanwhile, external model services are accelerating toward commercialization: On October 13, Physical Intelligence launched a beta API, enabling enterprises to upload scene data and leverage cloud-based computing power for fine-tuning and hosted inference; NVIDIA’s Isaac GR00T and Galaxy General’s π series—both embodied foundation models—are also advancing in parallel; furthermore, on October 23, Galaxy General and Unitree announced the joint development of Galbot ET1, covering the entire end-to-end pipeline from product definition to intelligent capabilities.
In this context, the division of R&D responsibilities among robotics companies is undergoing a restructuring: in-house development no longer automatically translates into a technological premium; instead, it must return to the core principles of business—using “full‑cycle cost accounting” and a “four‑stage return on investment framework” (R&D → pilot → replication → operations) to determine whether each investment effectively addresses unaddressed capability gaps, enhances delivery efficiency, and generates reusable assets. A controlled experiment demonstrating that Figure’s Index pre-training boosts the success rate of long‑range tasks from 9% to 56 serves as a quintessential example of value‑driven practice.